A modulo analog-to-digital converter (ADC) avoids clipping loud signals by folding them. Whenever the input crosses a threshold $\pm\lambda$, a feedback circuit subtracts a fixed voltage step, so the digitizer only ever sees a small range. Software then "unwraps" the folded samples back into the original large signal. The most robust unwrapping method is sequential linear prediction. It predicts the next sample from past unwrapped samples and picks the integer number of folds that makes the measurement agree with the prediction. This works exactly as long as the prediction error stays below $\lambda/2$. Real circuits break that assumption in a systematic way. Comparator thresholds and the reference that sets the fold step drift with temperature and age. Feedback delay makes the circuit fold late, and by different amounts on rising and falling edges. A single wrong integer decision then shifts every later sample by a full fold step, and the error never corrects itself.
The core idea of this note is that every fold is also a calibration measurement, and it costs nothing extra. If the fold step, the thresholds, or the delay are wrong, the unwrapped signal shows a small, repeatable jump at exactly the samples where folds happened. At all other samples the predictor sees only its ordinary error. Comparing prediction residuals at fold samples against residuals at slope-matched non-fold samples cancels the unknown signal. What remains is an estimate of the hardware's systematic parameters. This is the same logic by which reactor-neutrino experiments cancel unknown flux by taking ratios between near and far detectors. The estimate sharpens as the fold count grows. Large signals, which cause the most folds, therefore also supply the most calibration data.
Core Claim
We claim that a predictive unwrapper which jointly tracks the signal and a small vector of hardware systematics can keep integer fold decisions correct under realistic drift and asymmetric loop delay, with no reference injection, no resets, and no offline recalibration. The systematics are four quantities: fold-step error, threshold offset, and rising and falling delay. The unwrapper estimates them from fold-sample residuals, using non-fold samples as a control. The claim is falsifiable in two ways. The first test is whether the estimated drift tracks a known injected drift with error that shrinks roughly as $\sigma_e/\sqrt{N_f}$, where $\sigma_e$ is the one-step prediction error and $N_f$ the number of folds in the averaging window. That scaling holds if fold-sample prediction errors are approximately independent. The second test is whether long runs that make uncalibrated prediction fail within minutes instead run failure-free.
Why this gap is real
The lossless-above-Nyquist result for modulo folding in Above the Nyquist Rate, Modulo Folding Does Not Hurt assumes an ideal, instantaneous fold with a constant $\lambda$. So does the prediction-based ADC architecture of A Modulo-Based Architecture for Analog-to-Digital Conversion. Both supply the method we build on, and both leave out the systematics we target.
The hardware evidence that motivates the work comes from A Modulo Sampling Hardware Prototype and Reconstruction Algorithms Evaluation. That paper built a comparator, microcontroller, digital-to-analog converter (DAC) and 12-bit ADC loop. It found that linear prediction and a line-spectral method survive a 5% noise floor at 10× range extension, while the higher-order finite-difference method fails. It also reports that loop delay pushes folded outputs outside the ideal $[-1,1]$ V range. It does not model threshold or reference drift.
The hysteresis work in The Surprising Benefits of Hysteresis in Unlimited Sampling models hysteresis and fold transients explicitly and reaches 28× extension on a breadboard. Its parameters, however, are tuned once and assumed static. We differ in treating those parameters as drifting quantities that are estimated continuously from the data.
Two other lines of work handle model trouble differently. The blind vector-process converter in Blind Modulo Analog-to-Digital Conversion of Vector Processes learns signal statistics adaptively. When prediction fails it detects the overload and resets resolution, which discards range during exactly the transients one cares about. It also assumes ideal folding. Residual Recovery Algorithm for Modulo Sampling treats the fold-induced residual as an unknown to optimize, but assumes exact, stationary boundaries. Our object is different: we estimate the residual's systematic component, not the fold pattern itself.
The systematic-uncertainty literature supplies the estimator's logic. Model-independent short-baseline oscillations from reactor spectral ratios cancels an unknown reactor flux by comparing detectors at different baselines. New Measurement of Antineutrino Oscillation at Daya Bay holds detector-to-detector energy-scale variation to 0.2% using calibration sources, including naturally occurring spallation neutrons. Our fold events play the role of those naturally occurring calibration sources. Non-fold samples play the role of the near detector. Neither paper touches folded acquisition.
Mechanism
We write $y_n$ for the measured folded sample and $k_n$ for the decoded integer fold count. We write $\hat{x}_n$ for the one-step linear prediction and $\hat{s}_n$ for the local slope estimate from past unwrapped samples. The nominal fold step is $2\lambda_0$. The residual is then
$$ r_n = y_n + 2\lambda_0 k_n - \hat{x}_n = e_n + \Delta k_n\,\delta + \hat{s}_n\,\tau_{\pm}\,\mathbb{1}[\Delta k_n \neq 0] + c . $$
Here $e_n$ is ordinary prediction error. $\Delta k_n$ is the change in fold count at step $n$. $\delta$ is the step-size error. $\tau_{+}$ and $\tau_{-}$ are the effective delays on rising and falling folds. $c$ is a slow offset.
At non-fold samples only $e_n$ and $c$ appear. At fold samples, the delay term enters as slope times delay because a late fold lets the signal run on for $\tau$ seconds before the step is applied. We track $\theta=(\delta,\tau_+,\tau_-,c)$ with a Kalman filter whose process noise is a slow random walk.
The filter is fed differences between each fold sample's residual and residuals at nearby non-fold samples with matched $|\hat{s}_n|$. Slope matching matters because folds preferentially occur at high-slope moments, where $e_n$ is also larger. Without matching, signal-model error would leak into $\hat\theta$.
The corrected residual $r_n - \Delta k_n\hat\delta - \hat{s}_n\hat\tau_\pm - \hat c$ replaces $r_n$ in the integer decision. The effective decision margin therefore rises from $\lambda/2 - |\text{bias}|$ back toward $\lambda/2$ minus the estimation uncertainty. The same bookkeeping flags wrong fold decisions: a wrong decision appears as a persistent $\pm 2\lambda$ offset in the control residuals, and the filter repairs it retroactively instead of resetting.
Research Program
The work has three stages.
The first stage is simulation. We use the Gaussian bandlimited processes of the blind vector-process setup and the oversampling, threshold and signal-to-noise sweeps of the residual-recovery suite. We inject known drift profiles into $\lambda$ and the fold step: linear ramps, thermal step responses, and slow sinusoids. We also inject asymmetric delays and a DC offset. We measure how quickly $\hat\theta$ converges and at what drift rate each unwrapper first mis-decodes an integer.
The second stage is hardware. We rebuild a loop matching the Zhu et al. prototype and drive it with the same test signals: single sine, two-tone, frequency-shift keying and amplitude-shift keying, all at 10× range extension. We heat the comparator and reference board with a controlled source and record hours of data, logging an independent high-resolution reference channel as ground truth. We also rebuild a hysteresis testbed in the style of Bhandari et al. and turn its potentiometers mid-run, which creates abrupt, known parameter changes.
The third stage, secondary and clearly marked exploratory, applies the same estimator across detector elements in modulo tomography. Here each detector's threshold is estimated from its folds across projection angles. We simulate folding on real sinograms from 2DeteCT with per-element threshold non-uniformity. This is the regime that The Modulo Radon Transform explicitly leaves out by assuming uniform thresholds.
Baselines
The naive reference point is sequential linear prediction with nominal $\lambda_0$, zero delay, and a one-time calibration at power-up, as in the predictive ADC of Tabak et al. and the Zhu et al. hardware evaluation. We run it on identical recordings with the same predictor order and training window. Any improvement we claim must come from the systematics tracking, not a better predictor.
The strongest existing competitors on robustness are the two methods that survived real hardware in Zhu et al., linear prediction and the line-spectral method with dynamic programming. The hardware paper reports that linear prediction recovered a 3.5 kHz signal and the line-spectral method a 2.5 kHz signal at 10× the dynamic range limit, with "near-perfect reconstruction" under a 5% noise floor. No numerical reconstruction error is reported on Lacuna for either, so we will re-measure both on our recordings.
For adaptive robustness, we compare against the overload-detect-and-reset converter of Weiss et al. That paper reports unfolding error probability as low as 0.001% in simulation, but under ideal folding. We run it with the same prediction filter length and learning rate tuned on a held-out drift profile.
For optimization-based recovery, we compare against the residual-recovery algorithm. It reports −60 dB error at an oversampling factor of 10 on synthetic signals, versus an oversampling factor of at least 30 for finite differences. We run it on the same synthetic sweeps with drift added.
Two ablations isolate our key ingredient. The first estimates $\theta$ from fold residuals alone, without slope-matched control differencing. This tests whether the near/far cancellation is what removes signal leakage. The second drops the slope-times-delay term. This tests whether treating overshoot as deterministic rather than as outlier noise matters.
Beating the baselines means three things on the same drifted recordings and compute budget. First, zero integer mis-decodes over runs where naive prediction and the reset method fail. Second, unfolding error rate at least an order of magnitude below the best baseline at matched drift rate. Third, reconstruction error against the reference channel no worse than the baselines achieve on undrifted data.
Why not just…?
Why not just read the fold counter? In hardware like the Zhu et al. prototype, the microcontroller already counts folds, so integer decoding can be skipped entirely. This is a real threat, and it changes rather than removes the problem. With counted folds, a drifting DAC step produces an error of $k_n\delta$. That is a staircase distortion growing with amplitude, a systematic scale nonlinearity that the counter cannot see. Our estimator addresses exactly this case, because the same fold-sample jump measures $\delta$. We will include counter-readout reconstruction as a baseline on the hardware runs and report its staircase error with and without our step-size correction. If the uncorrected counter error is below the noise floor at realistic drift, the hardware half of the proposal narrows to counter-free designs such as the hysteresis testbed.
Why not just recalibrate periodically with a known reference? Injecting a reference signal costs acquisition time. It also perturbs the instrument, the same contamination problem precision-physics calibration faces. And between calibrations it cannot follow fast thermal transients. We compare at equal overhead by running reference calibration at intervals chosen so its dead time equals our filter's compute share. We then report which method keeps lower error between calibration points. If drift in practice is slow enough that hourly calibration suffices, our method's advantage reduces to eliminating dead time, and we will say so.
Why not just detect overload and reset, or add guard bands? The Weiss et al. reset does stop runaway error. But each reset drops dynamic range extension to zero, and resets cluster during high-amplitude transients, the very events modulo sensing is meant to capture. A fixed guard band that shrinks the decision region trades drift tolerance for a lower permissible prediction error. That in turn raises the sampling rate needed for a given bandwidth. Our experiments count the fraction of high-amplitude segments lost to resets under each drift profile. We also measure the oversampling a guard-band design needs to match our fold-error rate.
Decisive Checks
The first check is whether $\hat\theta$ tracks injected drift with error scaling like $\sigma_e/\sqrt{N_f}$. In particular, raising signal amplitude, and therefore folds per second, should tighten the estimate. If this scaling fails and the slope-matched control does not fix it, the cancellation idea is wrong.
The second check is whether, on long heated hardware runs, the calibrated unwrapper stays failure-free while uncalibrated prediction and the reset method do not.
The third check is whether the delay term measurably reduces the out-of-range overshoot outliers that Zhu et al. observed. If the estimated $\tau_+$ and $\tau_-$ do not differ, or do not predict overshoot size from slope, the asymmetric-delay model is unnecessary and should be dropped.
References
- Above the Nyquist Rate, Modulo Folding Does Not Hurt
- A Modulo-Based Architecture for Analog-to-Digital Conversion
- A Modulo Sampling Hardware Prototype and Reconstruction Algorithms Evaluation
- The Surprising Benefits of Hysteresis in Unlimited Sampling: Theory, Algorithms and Experiments
- Blind Modulo Analog-to-Digital Conversion of Vector Processes
- Residual Recovery Algorithm for Modulo Sampling
- Model-independent <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" altimg="si1.gif" overflow="scroll"><mml:msub><mml:mrow><mml:mover accent="true"><mml:mrow><mml:mi>ν</mml:mi></mml:mrow><mml:mrow><mml:mo stretchy="false">¯</mml:mo></mml:mrow></mml:mover></mml:mrow><mml:mrow><mml:mi>e</mml:mi></mml:mrow></mml:msub></mml:math> short-baseline oscillations from reactor spectral ratios
- New Measurement of Antineutrino Oscillation with the Full Detector Configuration at Daya Bay
- 2DeteCT - A large 2D expandable, trainable, experimental Computed Tomography dataset for machine learning
- The Modulo Radon Transform: Theory, Algorithms and Applications